CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-14 11:07:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: fix logic for clearing padding with -ngl 0 (#13320)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-05-05 22:32:13 +02:00
committed by GitHub
parent 233461f812
commit 9070365020
6 changed files with 33 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-cuda/mmvq.cu
View File

@@ -513,6 +513,16 @@ void ggml_cuda_mul_mat_vec_q(
const int32_t * ids_d = ids ? (const int32_t *) ids->data : nullptr;
float * dst_d = (float *) dst->data;
// If src0 is a temporary compute buffer, clear any potential padding.
if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
GGML_ASSERT(ggml_is_contiguous(src0));
const size_t size_data = ggml_nbytes(src0);
const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
if (size_alloc > size_data) {
CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
}
}
const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1);
{